key: cord-0727418-obevjmbi
authors: Tracey, Marlon; Plemmons, Alicia; Belasen, Ariel
title: Throwing caution to the wind: How hurricanes affect COVID‐19 spread
date: 2022-03-14
journal: Health Econ
DOI: 10.1002/hec.4499
sha: c919537b612400a049178e4e6d982b843cb65b8b
doc_id: 727418
cord_uid: obevjmbi

This study exploits the pathway of Hurricane Laura to assess its impact on the spread of COVID‐19. Using US hospital data on confirmed and suspected adult COVID‐19 cases, we find average daily cases per week rose by more than 12% primarily in tropical storm‐affected counties in subsequent weeks. We suspect the key mechanisms involve constraints on social distancing for two reasons. First, there is significant evidence of storm‐induced mobility. Second, lower income areas endured higher growth in hospital cases during the post‐hurricane period. These findings provide crucial insights for policy‐makers when designing natural disaster protocols to adjust for potential respiratory viral illnesses.

their counties experiencing substantial/high transmissions. 2 This was before vaccines were available and after state mitigation policies generally ended, except for mask mandates in public spaces (see Raifman et al., 2020) . Given this setting, we analyze differential trends in mobility and hospitalized/ICU COVID-19 cases between storm-affected and unaffected counties over 6 weeks (August 7 to September 17), allowing for heterogeneity based on wind-force and income levels.

We examine 405 hospital facilities in 199 counties across Arkansas, Louisiana, Mississippi, and Texas. We focus on the Eastern and Upper Gulf Coast counties of Texas, which are hurricane-prone and more closely resemble Arkansas and Louisiana in demographic makeup. The National Oceanic and Atmospheric Administration provides the path of Hurricane Laura's eye and surface wind field data defined at three wind speed thresholds: 39, 59, and 74 mph. Figure 1 shows Laura's path and wind field with at least 39 mph winds. It traveled northwards along western Louisiana and weakened to a tropical storm before crossing into Arkansas. Hospitals are considered affected by Laura if they are located in counties within the wind field; otherwise, they are unaffected. Our sample comprises roughly one-half of hospitals in 93 storm-affected counties.

Most hospitals must report daily COVID-19 capacity data to the U.S. Department of Health and Human Services (HHS) for Federal response planning. 3 Starting July 31, 2020, HHS released weekly (Friday to Thursday) counts of adults currently hospitalized or in ICU with confirmed or suspected COVID-19, as well as data on staffed beds. 4 But it redacted non-zero cases fewer than four for privacy concerns-we address this econometrically in Section 3. Since reporting days per week vary from 1 to 7, we compute average daily cases per week. Focusing on hospital cases is important as they are less discretionary, given preferential access to testing, and federally required to be reported. 5 Our event-study analysis of this data spans 3 weeks prior to Laura's landfall on August 27, since hospitalization data are unavailable from July 27 to August 6 for Louisiana. 6 The post-event period ends September 17 to avoid confounding with other proximate hurricanes (e.g., Sally on September 16). We only analyze hospitals with data pre-and post-Hurricane Laura.

F I G U R E 1 Hurricane Laura's path and wind field comprising at least 39 mph winds Our analyses rest on four additional data sources. Given no changes in state mitigation and reopening polices during our study period (see Raifman et al., 2020) , we use state-level testing data from The COVID Tracking Project (2021) to gauge states' responses/attitudes to the pandemic based on tests availability and outbreak severity. We also rely on anonymized smartphone location data, from SafeGraph and PlaceIQ, to assess differences in mobility between storm-affected and unaffected counties. The two companies track tens of millions of smartphones via GPS pings daily. We use SafeGraph's census-block-group-level counts of devices completely at home and aggregate it to the county level, as a percent of all devices in a county. Using PlaceIQ data, Couture et al. (2021) provides two indices: (i) the average number of devices encountered at commercial venues by devices in a given county on the same day (i.e., device exposure), and (ii) the share of devices in a county that were in any other county in the last 2 weeks (i.e., location exposure). For each measure, we compute a county's average daily mobility level per week (as per HHS reporting period). These data are commonly used to study social distancing behavior (e.g., Dave et al., 2021; Weill et al., 2020) . Finally, to check for effect heterogeneity across income groups, we define low-and high-income thresholds as the first and third quartiles of the 2019 county median household incomes in our study states (i.e., $39,840 and $53,423), using American Community Survey 5-year data. 7

We employ the following event-study framework to estimate Hurricane Laura's impact on a community's average daily adult COVID-19 cases that require a hospital ( * ).

We measure * using HHS-reported data on , that is, the average of daily cases in week admitted to hospital , located in county in state . However, we only know ≤ 3 when weekly case counts are between 0 and 4 due to HHS suppression. Moreover, = 0 may not mean a hospital service area had no moderate-to-severe cases. It may only indicate few such cases existed, but they sought care elsewhere (e.g., urgent care centers) or held out at home (e.g., deterred by medical costs, bed capacity, hurricane ruins). Thus, when hospitals report under four cases in a week (i.e., censored observations), we assume * ≤ 3 ; otherwise, we set * = . 8 Table 1 shows pre-hurricane weekly hospitalized and ICU cases less than four occurred at a rate of 31%-35% and about 50%, respectively. When facilities had at least 4 cases in a week, they averaged about 16-18 hospitalized and 9-10 ICU cases daily.

Equation (1) adjusts for static cross-county differences and secular trends in outcomes using county and week fixed effects ( and ), respectively. 9 Via , we also control for log number of staffed hospital beds and log number of tests conducted in state in week . Finally, the indicator equals to 1 if Laura impacted county , and 0 otherwise. Interacting it with week dummy allows for estimating , the effect of Laura when weeks away from the week of landfall. Given is non-negative, continuous, and right-skewed with mode at zero, we let * have an exponential density * = −1 − * ∕ , where = * . We then estimate a censored exponential regression by maximizing this log-likelihood: log = ∑ , ∉ log ( ) + ∑ , ∈ log (3) , where  is the set of censored observations and (⋅) is the exponential cumulative distribution function. Estimates of , when 0 , partially test that affected and unaffected counties share parallel trends in outcomes. But when 0 , they provide post-hurricane weekly effects. We also estimate two variants of Equation (1), where (i) we replace the left-hand side with (mobility) , take as log (tests), and apply ordinary lease squares; or (ii) we interact ⋅ ℎ with income or wind speed levels. 10 Table 2 presents estimated marginal effects ( ) of Hurricane Laura on log of expected daily cases requiring a hospital at various times from week of landfall. Prior to landfall, differences in trends are small and insignificant for either confirmed-only or suspected/confirmed COVID-19 cases. A week after landfall, however, column 1 shows higher average daily confirmed hospitalized cases by 12.6% ( 0.119 − 1) , with larger effects 2-3 weeks later (17%-26%). We find similar results for suspected/ confirmed cases (column 3). One might argue the use of hospital facilities increased due to sheltering needs by asymptomatic or mildly symptomatic people who would have otherwise stayed home. But such cases would not require ICU interventions. Columns 2 and 4 show the trend for ICU cases is similar to hospitalized cases, with notably larger effects within 2-3 weeks (i.e., over 35%). These estimates may be lower bounds since Laura arrived while states had mask mandates, which tend to slow case growth (Chernozhukov et al., 2021; Karaivanov et al., 2021) . Although our estimates reflect growth in hospital cases, they are comparable in magnitude to Dave et al.'s (2021) finding that the Sturgis Motorcycle Rally in Meade County, South Dakota raised COVID-19 cases by at least 31% in the county and its neighbors or by at least 12% in the state.

A potentially key reason Laura spread COVID-19 is that it induced mobility, especially where it is harder to mitigate the spread. To explore this, Table 3 provides marginal effects of Laura on three mobility measures. Within a week of Laura's arrival, for affected counties, the share of people staying home fell by 2.1 pp, and movements to crowded areas (device exposure) and other counties (location exposure) rose by 10.4% and 0.7 pp, consistent with late evacuations, (1), which adjust for county and week fixed effects, as well as hospital and testing capacities. All coefficients are expressed relative to the week of landfall. They are interpreted as percentage change when transformed to ( − 1) × 100 .

Standard errors in parentheses are clustered at the county level. Pre-trend χ 2 -test is a joint test for any pre-hurricane differential trends in outcomes. *, **, *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. returns, rescues, and recovery efforts. Analyzing mobility effects by wind speed and income level provides greater insight into the mechanism. Figure 2 (Top) shows counties affected by tropical storm-force winds (39-73 mph) spiked in device (crowd) exposure post-landfall. Meanwhile, people from hurricane-affected counties (74+ mph) primarily fled their homes for less-affected counties. Conceivably, evacuees, who shed the virus or are exposed to it, are a major source of COVID-19 spread to nearby counties (Price et al., 2020) . In fact, Table 4 (Panel 1) reports higher growth in daily hospital cases after landfall in areas affected by tropical storm-force winds, particularly when 59-73 mph (by at least 22.3%); and no significant impact in hurricane-affected areas. When we compare mobility effects across income groups (Figure 2 , Bottom), the increase in pre-hurricane mobility is somewhat larger for lower income counties, but all experienced a similar surge after landfall. We know these mobility data do not reveal the extent travelers took COVID-mitigating steps and thus their specific risk of catching or spreading the virus. Given the same mobility response, however, we expect travelers from high-income areas to better social distance (e.g., avoid public transit and foot traffic, or isolate at home) than low-income ones. Thus, if mobility under constrained social distancing is a mechanism for storm-induced COVID-19 spread, then the effect should be more pronounced in lower income areas. Table 4 (Panel 2) provides income-group-specific effects of Hurricane Laura on hospital cases. Specifically, we observe at least a 28% increase in daily hospital cases for low-income counties, and smaller magnitudes for higher income ones. These results support our expected mechanism, and reveal important distributional consequences of a dual emergency.

Hurricane Laura was the strongest natural event in the United States since the onset of the COVID-19 pandemic in early 2020, providing a unique opportunity to analyze their interaction. We show Laura contributed to COVID-19 spread by increasing adult hospital cases in storm-affected counties of Arkansas, Louisiana, Mississippi, and Texas, mainly the low-income areas. Our results suggest a possible mechanism: disaster mitigation efforts increased mobility in counties with limited capacity to social distance. As more coronavirus variants appear and climate change extends hurricane seasons, our findings reveal a need to redesign evacuation/sheltering plans to make compatible with infectious disease protocols. This may be warranted even with vaccine availability given the potential for breakthrough infections, waning vaccine efficacy, and the lower vaccine take-up in coastal states. TRACEY et al. 5 Completely at home Device exposure Location exposure Note: Income thresholds are $39,840 (lower quartile) and $53,423 (upper quartile). Coefficients are averages of the post-hurricane weekly marginal effects for each level of wind speed or income, adjusting for county and week fixed effects, as well as hospital and testing capacities. All coefficients expressed relative to the week of landfall. They are interpreted as percentage change when transformed to ( − 1) × 100 . Standard errors in parentheses are clustered at the county level. Pre-trend χ 2 -test is a joint test for any pre-hurricane differential trends in outcomes by wind speed or income level. *, **, *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. 

Mass gatherings contributed to early COVID-19 mortality: Evidence from US sports. Covid Economics: Vetted and Real-Time Papers

Causal impact of masks, policies, behavior on early COVID-19 pandemic in the US

The relationship between in-person voting and COVID-19: Evidence from the Wisconsin primary

Strong social distancing measures in the United States reduced the COVID-19 growth rate

JUE Insight: Measuring movement and social contact with smartphone data: A real-time application to COVID-19

The contagion externality of a superspreading event: The Sturgis Motorcycle Rally and COVID-19

Face masks, public policies and slowing the spread of COVID-19: Evidence from Canada

JUE Insight: College student travel contributed to local COVID-19 spread

Record-breaking Atlantic hurricane season draws to an end

Tropical cyclone report Hurricane Laura (AL 132020

A model of the spread of the COVID-19 pandemic during a hurricane in Virginia

COVID-19 US state policy database

Cascading risks of COVID-19 resurgence during an active 2020 Atlantic hurricane season

Mitigating the twin threats of climate-driven Atlantic hurricanes and COVID-19 transmission

Coronavirus (Covid-19) data in the United States

COVID-19 reported patient impact and hospital capacity by facility

Social distancing responses to COVID-19 emergency declarations strongly differentiated by income

The authors are responsible for disclosing all financial and personal relationships between themselves and others that might bias their work.

The data that support the findings of this study are available from SafeGraph. Restrictions apply to the availability of these data, which were used under license for this study. Data are available at https://www.safegraph.com/academics with the permission of SafeGraph.